# code
  import time
  import numpy as np
  import onnx
  import onnx.helper as oh
  import onnxruntime
  
  
  def model_type(itype):
      return oh.make_model(
          oh.make_graph(
              [oh.make_node("MatMul", ["X", "Y"], ["Z"])],
              "b",
              [
                  oh.make_tensor_value_info("X", itype, ["a", "a"]),
                  oh.make_tensor_value_info("Y", itype, ["a", "a"]),
              ],
              [oh.make_tensor_value_info("Z", itype, ["a", "a"])],
          ),
          ir_version=10,
          opset_imports=[oh.make_opsetid("", 18)],
      )
  
  
  sess16 = onnxruntime.InferenceSession(
      model_type(onnx.TensorProto.FLOAT16).SerializeToString(),
      providers=["CPUExecutionProvider"],
  )
  sess32 = onnxruntime.InferenceSession(
      model_type(onnx.TensorProto.FLOAT).SerializeToString(),
      providers=["CPUExecutionProvider"],
  )
  
  N = 20
  for size in [256, 512, 1024]:
  
      # float32
  
      f32 = dict(
          X=np.random.randn(size, size).astype(np.float32),
          Y=np.random.randn(size, size).astype(np.float32),
      )
      # warmup
      for i in range(10):
          sess32.run(None, f32)
  
      # measure
      begin = time.perf_counter()
      for i in range(20):
          sess32.run(None, f32)
      duration = time.perf_counter() - begin
      print(f"float32 size={size}, time={duration / N}s per iteration")
  
      # float16
  
      f16 = {k: v.astype(np.float16) for k, v in f32.items()}
      # warmup
      for i in range(10):
          sess16.run(None, f16)
  
      # measure
      begin = time.perf_counter()
      for i in range(N):
          sess16.run(None, f16)
      duration = time.perf_counter() - begin
      print(f"float16 size={size}, time={duration / N}s per iteration")
  ```

</details>